Reproduce “Trade-off between Gradient Measurement Efficiency and Expressivity in Deep Quantum Neural Networks” on OpenResearch

About this paper

Koki Chinzei, Shinichiro Yamano, Quoc Hoan Tran, Yasuhiro Endo, Hirotaka Oshima

Chinzei et al. established a fundamental trade-off between gradient measurement efficiency and expressivity in deep quantum neural networks, proving that more expressive networks incur higher measurement costs. The authors introduced the Stabilizer-Logical Product Ansatz (SLPA), a quantum neural network architecture that achieves this theoretical efficiency limit, reducing the total measurement shots needed for training by factors like eight for certain tasks while maintaining performance.

Key insights

• A fundamental trade-off exists where higher expressivity in deep QNNs necessitates greater measurement costs for gradient estimation, and conversely, constrained expressivity can enable more efficient gradient measurement.
• Explicitly encoding problem symmetries into QNN architectures, as demonstrated by the SLPA, acts as a powerful inductive bias, leading to substantially improved gradient measurement efficiency and often better generalization.
• The Stabilizer-Logical Product Ansatz (SLPA) provides an optimal QNN design that reaches the upper bound of the derived trade-off, demonstrating a practical approach to construct efficiently trainable deep QNNs.

Summary by alphaXiv.